What Is Consumer Credit Risk?

Michael Whitman · Trading since 2017, full-time since 2020 · YouTube creator, 137K subscribers · Published 2026-09-14 · Updated 2026-10-03

A bank lending 50 million to a company writes a document about that company. A bank lending 50 million across 10,000 credit cards writes a model. The difference in approach drives everything else.

How it works

The exposures are numerous and small. No individual account is large enough to matter on its own, and no lender can economically investigate each one.

So the question is not whether this borrower repays. It is what proportion of borrowers with these characteristics repay, which is a question data can answer.

Scale is what makes this work. A hundred loans behave unpredictably; a hundred thousand behave close to their expected rate, which is the same principle insurance runs on.

The correlated part

Individual defaults are close to independent in normal conditions. One person losing their job tells you nothing about the next borrower.

In a recession they stop being independent. Unemployment rises across a region, and every borrower in that region becomes more likely to default at the same time.

So the whole loss distribution shifts. A portfolio modeled on independence badly understates what happens in the one scenario the capital was being held for.

A worked example

A card portfolio of 100,000 accounts with an expected annual loss rate of 4%, priced at an interest rate that covers that plus funding, costs and a margin.

In a normal year the outcome lands near 4%. The statistical approach works, the business is profitable, and no individual account decision was ever revisited.

Unemployment rises two percentage points. The loss rate goes to 8% — not because the model was wrong about individuals, but because the common factor it treated as stable moved.

At 8% the portfolio is loss-making, and the lender cannot reprice existing balances fast enough to respond. That gap between the modeled rate and the stressed rate is the entire risk of the business.

What the models actually use

Payment history dominates. Past behavior on credit obligations is the strongest available predictor, and it is why a credit score weights it most heavily.

Utilization matters more than most people expect. How much of an available limit is drawn is a strong signal, which is why a fully used card damages a score even when payments are current.

Stability of information helps. Length of history, consistency of address and employment all contribute, which systematically disadvantages people who are young, recently arrived, or have simply never borrowed.

And affordability is assessed separately. Regulation in most jurisdictions now requires lenders to check that a borrower can afford the repayments, not only that they are statistically likely to make them — a distinction that came directly out of the last credit cycle.

The original data

Measured on the S&P 500 since 1950, 93% of trading days ended below the previous record, the worst fall was 56.8%, from 2007 to 2009, and the longest below-peak stretch ran 7.5 years, from 1973 to 1980, before new records followed.

That shape — long periods of grinding pressure rather than single shocks — is what a credit downturn looks like on a lending book. Losses do not arrive in one day; they accumulate over quarters as arrears roll through to write-off.

And this site’s fee measurement shows the compounding logic from the borrower’s side: 150 basis points consumes 36.5% of a thirty-year balance. Consumer credit rates are an order of magnitude above that, which is what credit card interest works through.

Why it matters to somebody who is not a lender

It explains the pricing you are offered. A rate is not a judgment about you personally; it is the expected loss rate of the group the model has placed you in, plus costs and margin.

It explains why scores behave oddly. Closing an old account or paying off a card entirely can lower a score, because the model measures predictive characteristics rather than virtue.

And it explains credit availability through a cycle. Lending tightens sharply in downturns because the models reprice the common factor, not because individual applicants changed.

Which makes the whole system procyclical. Credit is most available when it is least needed and withdrawn when it is most needed, and that is a property of statistical lending rather than a failure of it.

What regulation now requires

Affordability assessment. Most jurisdictions require lenders to check that repayments are sustainable from the borrower’s income, not merely that the model predicts repayment.

Explainability. A decline based on a model must be explicable, which constrains how opaque the scorecards may be and gives applicants a right to the reasons behind a decision.

And protected characteristics may not be used. Models cannot lawfully price on race, sex or similar attributes, which also requires testing for proxies that correlate with them.

That last requirement is harder than it looks. Postcode correlates with a great deal, so a model using it can reproduce a prohibited outcome without ever naming the prohibited variable - which is why fairness testing is now a standing part of building these systems rather than a one-off check.

How this connects

It is credit risk with the lens reversed. Corporate lending studies one borrower’s business risk; consumer lending studies a population and prices its average.

Mortgages are the largest consumer exposure and behave differently from cards. A mortgage is secured on a home, so losses depend on house prices as well as on employment, and both can fall together.

The debt service coverage ratio is the business equivalent of an affordability check, comparing income with repayments in the way regulators now require for households.

Lenders rarely keep all of it. Pools of card, auto and home loans are packaged and sold into the bond market, which spreads losses to investors and turned the correlated defaults described above into a channel for systemic risk.

The modeling trap is overfitting. A scorecard tuned on benign years performs well in a backtest on those same years, which is why stress scenarios have to be built from outside the data.

For the lender, a recession is tail risk: rare in the sample and dominant in the losses.

From the household side the defences are unglamorous. An emergency fund stops a lost income becoming a missed payment, a written budget keeps utilization down, and the debt snowball clears balances in an order that is easy to stick to.

And rising inflation squeezes the same borrowers from the other side, lifting living costs faster than many incomes adjust.

When it fails

The characteristic failure is modeling on a period without a recession. A lender builds its scorecards on several years of benign data, the models fit beautifully, and expansion into thinner credit looks justified by the observed loss rates. The data never contained a downturn, so the correlation between borrowers was never visible in it. When unemployment moves, every account in the book responds to the same factor at the same time, and the loss rate goes to a multiple of anything in the training sample. The model was accurate about the world it was shown.

A second failure is treating defaults as independent, which they are until they are not.

A third is expanding into weaker credit late in a cycle, when the observed loss rate is at its most flattering.

A fourth is confusing affordability with willingness to pay. They are different questions and both matter.

The question: Losses at a record low. Lend more?

And a fifth is assuming a score measures a person. It measures a group’s historical behavior and applies it, which is useful, statistical, and not the same thing.

Related

Credit score covers the number these models produce. Credit card interest covers what the risk is priced at. And concentration risk covers the common factor that makes a diversified book behave as one.

This page is educational, not financial advice. Test every idea on your own charts before risking money.